Lesion-Aware Graph-Augmented Deep Framework for Pancreatic Cancer Detection from CT

Because pancreatic cancer lesions are tiny, iso-attenuating, and partially obscured by surrounding arteries and ducts, early detection is still challenging. The end-to-end pipeline proposed in this study includes lesion-aware feature fusion, hybrid U-Net tumour masking, radiomic texture mining, physics-aware pre-processing, a MobileViT backbone with a lightweight graph head, and post-hoc probability calibration. It makes use of the 1,418 high-resolution DICOM slices (typically 512×512) from the Pancreatic CT Images collection. Slice-level performance is 0.942 for accuracy, 0.936 for macro-F1, 0.972 for AUROC, and 0.969 for AUPRC, according to the held-out test set. The accuracy increases to 0.960, the macro-F1 increases to 0.957, the AUROC increases to 0.986, and the AUPRC increases to 0.983 when the data is aggregated at the patient level. The Expected Calibration Error decreases from 0.031 to 0.009 when the temperature is scaled, improving the accuracy of risk assessments. With a Dice score of 0.82±0.09 (median 0.84; HD95 7.3 mm), the segmentation branch is likely to get worse as the tumours get smaller than a centimetre. Lesion-aware attention yields the greatest gains in calibration and discrimination, according to ablations (ΔAUROC −1.4 points, ΔECE +0.012 upon removal). There are also some advantages to radiomics (-0.3 points) and graph (-0.7 points). The site shift is minimal (worst site: AUROC 0.958), and the AUROC decreases steadily under robustness stress (motion S3: 0.940; noise S3: 0.945; bias-field S3: 0.949). The sensitivity and specificity of a Youden-optimal threshold are 0.927 and 0.935, respectively. The sensitivity and specificity of a screening operating point are 0.965 and 0.881, respectively. With throughput (about 28 ms each slice; 1.48 s per study) and 98.5% coverage following quality gates, the system can be used in real-world scenarios. According to the findings, pancreatic cancer can be detected on CT images using calibrated, lesion-aware, graph-augmented fusion that is efficient and low in processing overhead.